Earned Context
A reviewer that starts from zero every time is doing the same expensive work over and over. Sigilix doesn’t. Each review leaves something behind — a verified layer of understanding about your repository — and the next review, the CLI, and the chat all build on it instead of rediscovering it. This is the engine that fuels Sigilix’s models: verified repository context, deposited by the work and fed back in on every run, so our tuned models reason about your system rather than software in general.
This is what makes Sigilix judge a diff in context. Before any specialist runs, Sigilix expands the change against the whole repo: who calls the function being modified, what depends on it, which symbols it touches. The diff is read against the repository, not through a keyhole.
Context before judgment
A finding like “this function no longer handles the null case” is only correct if you know every caller. A window of a few diff lines cannot tell you that. Sigilix pulls in callers, dependencies, and symbols from the code graph and index, and reconciles against review memory, so a specialist is reasoning about the real blast radius of a change — not just the lines that changed.
The layers a review deposits
Index
A searchable, semantically-embedded map of the repository — files, symbols, and structure — so context can be retrieved by meaning without re-reading the whole tree each time.
Code graph
Symbols plus import and call edges, kept per repo and per branch. This is how Sigilix finds the blast radius of a change and judges the diff against the whole repo — see The Code Graph & Blast Radius.
Trust ledger
The record behind proof-tier receipts — what was verified, reproduced, and grounded, and how findings earned their tier over time. See Verification & the Trust Ledger.
Review memory
What Sigilix has learned about this repo: taught rules (see Conversational Learnings) and the statistical accept/dismiss signal.
Evidence manifests
The cited code, anchors, and checks that back each finding — the provenance that lets a claim be re-examined rather than re-derived.
Each layer is verified understanding, not a cache of raw text. It is built by the same gates that decide what posts, so the context a later review inherits is context that already earned its place.
This injected context is honored most reliably by Sigilix’s own hosted models, which are tuned to prioritize it over their general knowledge. A model brought via bring-your-own-key may not use the earned context as fully — see the grounding trade-off there.
Who draws on it
The earned-context layer is not just for the next PR review. It is the substrate the rest of the product reads from:
Review parity in the terminal, drawing on the same index, graph, and memory the hosted reviews built.
Grounded answers about the repo and its review history, because the history is a retrievable layer.
CI-failure and issue triage reason against the graph and memory to produce grounded root causes.
Why it compounds
The first review on a repo pays to build the index and graph. Every review after that adds to the trust ledger, sharpens review memory, and accumulates evidence manifests. The understanding deepens with use — which is why a reviewer that remembers your repo gets more useful the longer it works on it, and why the same understanding can power the CLI and chat without rebuilding it from scratch.
Read next
The mechanics of the graph and index — edges, versioning, and blast-radius analysis.
The checks and receipts that make each layer verified rather than cached.
The gates that build the trust ledger and evidence manifests as findings earn their place.
How the terminal and Deep-Research Chat read from the earned-context layer.